The extent and magnitude of bias in case-crossover studies of real-world non-transient medications patterns: a simulation study with real-world examples

Background: A case-crossover study is a self-controlled design most appropriate for evaluating transient medication exposures but it has increasingly been used in studies of chronic medications and can cause bias in effect estimates. Previous studies have demonstrated the presence of persistent user bias in the case-crossover design where they examined persistent pattern with the assumption that once patients begin a chronic medication, they will continue to use that medication. However, we know that there are significant variability in the patterns of chronic medication use instead of one persistent pattern. Therefore, the magnitude of the bias is unclear and may vary based on the different patterns of chronic medication use when using a case-crossover design.

Objectives: The goal of this study was to evaluate the magnitude of this bias across different medication use patterns using a simulation study and a real-world example of Sodium-Glucose Cotransporter-2 Inhibitors (SGLT2) users with both a positive (genital infections) and negative (retinal detachment) outcomes.

Methods: We conducted a simulation study evaluating the bias introduced by 1) 7 common medication patterns separately, and 2) cohort with combined 15 different patterns. We evaluated each scenario under risk ratios of 0.5, 0.75, 1.00, 1.5, and 2.0. Each approach was analyzed using conditional logistic regression comparing the probability of exposure on the outcome day to 30 days prior. A case-time-control design was used in each of the scenarios. Sensitivity analysis was performed to evaluate the impact on the estimates when changing the length of risk and control windows from one day to 30 days. We conducted a real-world example focusing on SGLT2 users as real-world examples.

Results: The case-crossover design resulted in unbiased estimates when patterns were consistent with transient exposures but were biased upward with prolonged exposure patterns. The magnitude of the bias varies by patterns or pattern combinations. When evaluating prolonged exposures individually or combined as a cohort with mixture patterns, case-time-control with extended risk and control window (30 days) produced unbiased results (mean bias≤0.03). The real-world examples findings were consistent with the simulation work.

Conclusions: In this evaluation, we found that the case-crossover design produces true estimates with patterns that represent transient exposures, and upward biased results with non-transient medication patterns. Researchers that used case-crossover design to evaluate non-transient exposures should implement recommended methods such as case-time-control with extended risk and control windows to account for biases.